Subscribe
Anthropic decided to add invisible watermarks to all outputs from its latest Claude models mainly because of European regulation, not just technical reasons or a new focus on transparency. Since the company cannot easily limit watermarking to certain regions, the EU’s transparency rules under the AI Act now affect Claude worldwide, including APIs, consumer products, coding tools, and cloud partners. What started as a local rule has become a global standard.
This is an example of the Brussels Effect. The European Union sets rules, enforces penalties, and takes advantage of the difficulty in maintaining different product versions to spread its standards worldwide. Brussels officials call this strategic leadership. By being the first to mark AI-generated content, Europe wants to set the global standard for “responsible” AI and encourage others to follow. In reality, though, Europe is now seeing if being first with regulation still brings influence or just adds costs that competitors without these rules can avoid.
The First-Mover Illusion
This idea is well known. The large EU market and the risk of big fines push companies to raise their standards. Once companies set up compliance systems for Europe, it is usually easier to use them everywhere instead of running separate systems. GDPR showed that this approach can work. The AI Act’s transparency rules, including the Code of Practice for marking synthetic content, aim for similar results.
But AI is different from personal data. Model weights and inference pipelines can be changed more easily. Providers can use the same main model and just add a watermark at the end. Open-weight models and providers outside the EU do not have to follow these rules. Soon after Anthropic’s announcement, open-source tools came out that can weaken or remove watermarks. The difference is clear: closed, regulated systems show the mark, but many other options do not.
EU policymakers think that the benefits of following the rules will eventually be greater than the inconvenience. They say transparency will become normal and that platforms, businesses, and governments will want marked content. They expect users who care about privacy to adjust. While this has happened in some areas before, early signs in generative AI suggest things may turn out differently.
Short-Term Pain, Concentrated Costs
At first, the costs mainly affect companies that follow European rules and users who want unmarked output. Anthropic admits that the watermark is added at the model level and can stay even after small edits. It shows Claude was involved, but not that it wrote everything. For many professionals, academics, and creatives, this difference does not help much. A visible mark on a lightly edited draft, translation, or reviewed code can create a stigma that some people will want to avoid.
Users have already voiced this complaint. Many see the watermark as a digital scarlet letter, a lasting sign that AI was involved, even if most of the work was done by a person. In fields where originality, privacy, or clear human authorship are important, people are likely to choose other options. Open-weight models and providers that do not mark their output are becoming more attractive. Some business clients worried about being detected are now looking for alternatives that offer more privacy. This competitive disadvantage is not just a theory; it is already showing up in user feedback and early behavior.
This is a real example of competitive disadvantage. Europe’s strict early rule makes compliant systems more expensive than non-compliant ones. Companies that follow the European standard are less competitive in the short term for many users. People who value speed, flexibility, and control are more likely to take their work elsewhere. Europe pays the regulatory costs, while others benefit from unmarked output.
Europe has responded by relaxing some requirements instead of making them stricter. Deadlines for high-risk rules have been pushed back, efforts to simplify regulations are ongoing, and some industries have been given exceptions. These changes show a key issue: supporters of regulatory leadership also see that too much bureaucracy could limit Europe’s access to the technology it wants to oversee. Worries about competitiveness are already shaping the first-mover strategy.
What Europe Gets Wrong About Power
A bigger mistake is thinking that making rules means having control over technology development. In areas where Europe does not have top labs, strong computing resources, or major private investment, being first with regulations is less powerful than it used to be. The United States still leads in advanced model development and computing power, while China follows its own path. Open-source communities respond to incentives that European laws cannot reach.
When regulated companies apply European rules worldwide, but unregulated options remain available, the market does not automatically follow the European standard. Instead, the market splits, and users who care about privacy, speed, or cost move away from marked systems. If marked systems become a smaller or slower part of the ecosystem, Europe’s ‘leadership’ ends up creating rules that many innovative players avoid. A targeted and limited approach could address these issues such as public-interest deepfakes, major disinformation campaigns, or significant automated decisions, without imposing universal rules. Europe instead adopted a broad policy, reflecting a preference for control over experimentation. This choice incurs costs that are not distributed equally.
Competitiveness Requires More Than Rules
Long-term competitive advantage in AI depends on attracting talent, securing major investments, leveraging strong energy and computing resources, and enabling rapid model improvement. Clear rules can help make a place more attractive, but rules that are too strict and hurt local and foreign companies do not. Europe’s current approach could slow progress and lacks the reach or technical strength to make its standards unavoidable.
The Anthropic watermark is a clear example. A regional transparency rule, applied worldwide because geofencing is not practical, quickly pushes users who want privacy to look for other options. Europe is betting that this will only be temporary and that unmarked output will eventually be rare. This bet could work if other major regions follow or if detection becomes common and expected. It could fail if open systems and non-European providers attract users who want more flexibility and privacy.
In the short term, the situation is clear. Companies that follow the European model are less appealing to many users. Europe is pushing its rules abroad but paying the price in competitiveness. Here, being first seems more like a self-imposed limit than real strategic leadership. In the end, the market will decide if this approach was right.